acceptodds
Under review as a conference paper at ICLR 2027

Defending Multilingual Refusal Capability Against Three-Layer Cascading Attacks with Behavior-Invariant LoRA Distillation

Abstract

Safety-aligned language models often encode refusal behavior in internal activation directions, but these representations can be fragile across layers, prompt positions, and languages. We introduce Three-Layer Cascading Attack (TLCA), a representation-guided attack that produces ordinary textual prefixes and suffixes. Once constructed, the attack requires neither access to model internals nor any modification of model parameters, activations, or decoding procedures. We then propose Behavior-Invariant LoRA Distillation (BILD), which freezes the base model and trains low-rank adapters to reproduce its clean behavior under adversarial perturbations. We show that a fixed-template LoRA adversarial-training baseline can suppress harmful compliance under attack but also causes severe benign-utility degradation and repetitive refusal behavior. In contrast, BILD distills the model's own clean refusals and helpful answers, thereby avoiding universal-refusal collapse while preserving instruction-specific response diversity in multilingual contexts. We additionally compare BILD with Refusal Feature Adversarial Training (ReFAT), a representation-level adversarial-training method, under the same TLCA evaluation protocol. Extensive multilingual experiments show that TLCA substantially weakens refusal behavior across model families, whereas BILD consistently restores robustness while maintaining strong compliance on harmless requests.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multilingual refusal capability, three-layer cascading attack, behavior-invariant LoRA distillation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rTy5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies multilingual refusal robustness through an attack–defense framework. The paper first introduces TLCA, which uses refusal-related representations at three layers to construct textual adversarial prefixes and suffixes. It then proposes BILD, which trains LoRA adapters to reproduce the original model's clean responses on both clean and attacked harmful/harmless prompts. The authors report reduced harmful compliance under TLCA while maintaining high non-refusal rates on benign requests.

Critical strengths

The paper studies multilingual refusal robustness through an attack–defense framework. The paper first introduces TLCA, which uses refusal-related representations at three layers to construct textual adversarial prefixes and suffixes. It then proposes BILD, which trains LoRA adapters to reproduce the original model's clean responses on both clean and attacked harmful/harmless prompts. The authors report reduced harmful compliance under TLCA while maintaining high non-refusal rates on benign requests.

Critical weaknesses

I have significant concerns about the presentation of the paper. As a human reader, I find it very hard to read. The abstract and the introduction introduce several terms and claims without sufficiently clear explanations. It also seems that the authors do not really pay attention to some content generated by LLM. For example, in line 67, the authors mention "a method introduced at the International Conference on Learning Representations (ICLR) 2025".

Questions for authors
  • How well does the refusal detector agree with human judgments across languages, particularly for responses that contain refusal phrases but still provide harmful information?

  • Why are the original models’ harmless compliance rates omitted from Table 2? Also, why in Table 3 are the results for Chinese Harmless N/A?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer

We thank the reviewer for the careful reading, and we apologize that the presentation made the paper difficult to follow. We have thoroughly revised the writing. For the two questions, one led us to add a new manual audit, and the other concerns content that was already in the paper but is now stated more explicitly.

W1. Presentation quality and LLM-generated phrasing.

We took this concern seriously and revised the manuscript throughout. Concretely:

  • The sentence flagged at line 67 ("a method introduced at the International Conference on Learning Representations (ICLR) 2025") has been removed; the work is now cited conventionally.
  • The abstract and introduction were rewritten: all technical terms and claims therein are now clearly explained at first use.
  • A full-text language pass removed rhetorical and non-idiomatic phrasing, and every number quoted in the text was re-verified against the tables.
  • An AI Use Statement was already included (page 51): it discloses that generative AI tools were used only for grammar correction and minor language edits; all content was reviewed and verified by the authors, who take full responsibility for it.

If any specific passage remains unclear, we would be grateful for a pointer and will clarify it promptly.

Q1. "How well does the refusal detector agree with human judgments across languages, particularly for responses that contain refusal phrases but still provide harmful information?"

We added a manual audit (Appendix G, Tables 10–11). We randomly sampled 60 responses per language and condition — 1,680 responses in total (14 languages × 2 conditions: the original model under TLCA, and BILD under TLCA; Qwen3-8B outputs). The annotator assigned refusal/non-refusal labels without access to the detector's predictions, and responses were translated with Google Translate before annotation to enable consistent inspection across languages.

The detector and the human annotator assigned the same label for 97.4% of the 1,680 audited responses. The audit also characterizes the error in both directions. In one direction — the failure mode the reviewer asks about — 17 responses (9 from the original model, 8 from BILD) contained a refusal phrase yet still provided harmful information; for these the detector counts a refusal, slightly underestimating harmful compliance. In the other direction, some genuinely refusing responses avoided any dictionary phrase and were counted as non-refusals, slightly overestimating harmful compliance. Phrase-based evaluation therefore has bidirectional error, and the two directions partially cancel. Given the 97.4% agreement observed in the manual audit, the residual error is small and acceptable, which supports the use of the dictionary-based detector for this evaluation. The audit data and both error directions are now reported explicitly in Appendix G.

Q2. "Why are the original models' harmless compliance rates omitted from Table 2? Also, why in Table 3 are the results for Chinese Harmless N/A?"

  • We initially reported only the defended models' harmless compliance in Table 2, on the reasoning that BILD's absolute harmless rates (96.0%, both without attack and with the TLCA prefix–suffix added) already bounded any possible loss of benign utility to within a few points. The reviewer's question convinced us that a direct comparison is more convincing than a bound: we have now run the original models on the same harmless test split and added them to Table 2 (94.0% clean, 96.4% under TLCA), which shows that BILD matches the original model on benign requests.

Factual Error: The Chinese harmless N/A was already explained in Section 4.1

The N/A entry is a property of the PolyRefuse dataset, not an experimental choice or an unexplained omission: the PolyRefuse test split does not contain Chinese harmless instructions. This was already stated in Section 4.1 of the original submission: "PolyRefuse does not provide Chinese harmless instructions in the test split, so harmless compliance is reported for the other thirteen languages but not for Chinese." In the revision we have additionally placed the same note directly beneath Table 3, so the N/A is now also explained at the point where the reader encounters it. Harmless compliance is reported for the other thirteen languages.

We hope the revised presentation and the new audit address the reviewer's concerns.

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer rTy5,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 2 Days

Dear reviewer rTy5,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 2 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in One Day]

Dear reviewer rTy5,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only one day, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Reviewer_XqQKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the refusal capabilities of multilingual large language models and their robustness against jailbreak attacks. It proposes a Three-Layer Cascade Attack (TLCA), which constructs adversarial prefixes and suffixes by optimizing refusal-related representations across multiple model layers to weaken the models' safety refusal capabilities. To defend against this attack, the authors further propose Behavior-Invariant LoRA Distillation (BILD), which uses the original model's normal responses to harmful and benign requests as supervision signals to preserve the model's original behavior under adversarial perturbations.

Critical strengths
  1. The research problem is practically relevant. The paper focuses on the vulnerabilities of model safety refusal capabilities in multilingual settings, considering both robustness against attacks and responsiveness to benign requests.

  2. The experimental evaluation is relatively extensive. The experiments cover five different models and 14 languages. The authors also compare their method with ReFAT, demonstrating the effectiveness of the proposed defense under the fixed TLCA attack setting.

Critical weaknesses
  1. Insufficient comparison with existing attack methods and limited evaluation of defense generalization. There are already many studies on representation-guided jailbreak attacks and adversarial prefix/suffix optimization, but the authors do not compare TLCA with any existing attack baselines. Moreover, the proposed defense is trained specifically against this attack configuration, raising the possibility that it merely learns to ignore certain tokens at the beginning and end of the input. The paper also does not evaluate whether the proposed defense leads to performance degradation on the model's original capabilities.

  2. The comparison between ReFAT and BILD appears unfair. BILD uses a significantly larger LoRA rank, while the training compute and number of training samples also differ. These differences may affect the validity of the comparison.

  3. The paper lacks sufficient implementation analysis. For example, why is a three-layer design necessary for TLCA?

Questions for authors

See weaknesses.

8 comments
AuthorsOpenReview ↗

Reply to Reviewer

We thank the reviewer for the detailed and technically substantive review. We address each point below; some points led us to run new experiments, while others concern analyses that were already part of the paper but may not have been presented clearly enough, which we clarify here.

W1(a). "Insufficient comparison with existing attack methods."

We added the two canonical optimized-text attack families as baselines: a GCG-style universal suffix attack (20-token shared suffix, 500 coordinate-gradient steps, top-256 per position, 512 candidates per step; Appendix C) and an AutoDAN-style evolved readable template (Appendix D). Macro-averaged over fourteen languages (Table 1):

AttackHarmful compliance (original model)Gain over clean (pp)
TLCA72.3%+45.7
GCG38.0%+11.4
AutoDAN44.9%+18.3

TLCA is the strongest of the three attacks for every evaluated model. These results position TLCA against existing attack methods quantitatively.

W1(b). "On the generalization of the defense." BILD is trained only on TLCA-perturbed inputs and never sees GCG or AutoDAN during training. We therefore evaluate whether the defense transfers to these two structurally different attack families (Table 1):

Attack (unseen in training)OriginalBILDΔ (pp)
GCG38.0%21.8%−16.2
AutoDAN44.9%18.8%−26.1

BILD reduces harmful compliance under both attacks, and Δ\DeltaBILD is negative in every model–attack row of Table 1 (all five models, both attack families). This indicates that the defense does not merely memorize the TLCA wrapper but transfers to attacks it was not trained against.

W1(c). "The defense may merely learn to ignore certain tokens at the beginning and end of the input."

The AutoDAN result alone rules out this interpretation. AutoDAN does not place adversarial tokens at fixed prefix/suffix positions: it embeds the instruction inside a readable template, so adversarial content appears both before and after the instruction and is interleaved with it (Appendix D). There is no separable prefix--suffix pair for the model to "ignore". If BILD had merely learned to disregard tokens at the beginning and end of the input, it should fail completely under AutoDAN; instead it reduces harmful compliance from 44.9% to 18.8% without having seen AutoDAN during training (Table 1). GCG provides the complementary case, where the adversarial content is a suffix only and consists of non-readable token strings, and BILD reduces harmful compliance there as well (38.0% →\rightarrow 21.8%).

Second, the harmless-instruction control speaks against an ignore/refuse rule tied to the attack wrapper. BILD's training mixture wraps the same prefix–suffix around harmless instructions and requires helpful responses. The main evaluation measures this case for the defended model: we take the benign (harmless) test instructions, prepend/append the TLCA attack prefix and suffix to them, and record how often the model still answers helpfully. A position- or string-based ignore/refuse rule would destroy benign utility, yet BILD still answers 96.0% of these TLCA-wrapped harmless instructions helpfully; we have now additionally measured the original model under the same TLCA-wrapped harmless input, which answers 96.4% (Table 2).

The response-diversity analysis (Appendix I) further shows that generations remain prompt-specific rather than template-driven.

W1(d). "The paper does not evaluate whether the defense leads to performance degradation on the model's original capabilities."

The capability most directly at risk from refusal-oriented training is benign utility. BILD's own benign utility was part of the main evaluation: on the complete harmless test split, BILD answers 96.0% of instructions helpfully both without attack and with the TLCA prefix and suffix added. Since this absolute rate is within a few points of typical original-model behavior, any damage to benign utility was already bounded regardless of the exact original-model baseline, which is why that baseline was not included initially. For rigor, we have now additionally measured the original models on the same harmless test split: 94.0% on clean inputs and 96.4% with the TLCA prefix–suffix added — confirming that BILD's 96.0% in both conditions matches the original model, i.e., no degradation is observed. Unique-response counts (Appendix I) confirm that response diversity is also preserved. Methodologically, BILD modifies behavior through only 823–972 prompt-specific distillation samples per model (Table 7) with the backbone frozen, which limits the scope for capability drift. We agree that general capability benchmarks (e.g., MMLU-style suites) would be a useful complement, and we note this as future work; our evaluation targets the utility dimension that this defense family could plausibly degrade.

W2. "The comparison between ReFAT and BILD appears unfair: BILD uses a significantly larger LoRA rank, while the training compute and number of training samples also differ."

We clarify why the configurations differed in the first place, and we then report a controlled experiment that removes this difference.

Why the original configurations differ. The differing LoRA rank, training-set size, and compute were not an arbitrary choice on our part but a consequence of faithfully reproducing each method as published. For the ReFAT baseline we followed the original ReFAT paper's own method and hyperparameter settings exactly, rather than re-tuning it to match BILD. BILD's configuration, in turn, is the one our method requires (rank-64 LoRA, etc.). The two configurations therefore reflect each method's authors' recommended settings, not a deliberate asymmetry introduced to favor BILD.

Controlled experiment with aligned configurations. To isolate the method from the configuration, we re-trained ReFAT with every training-side variable aligned to BILD (Appendix E.1, Table 9). On Llama-3.1-8B we set LoRA rank 64, scaling factor 128, dropout, target modules, training-set size, epochs, learning rate, batch size, gradient accumulation, sequence length, gradient clipping, learning-rate schedule, warmup ratio, and weight decay all to BILD's values. Only the ReFAT-specific mechanism (refusal-feature estimation and ablation) was retained, so the experiment isolates the method from the configuration.

Aligning the configuration changes ReFAT's results only marginally and in mixed directions (Llama-3.1-8B, Table 9):

Metric (Llama-3.1-8B)ReFAT (original)ReFAT (aligned)BILD
TLCA-attacked harmful (↓)55.9%59.0%12.3%
Clean harmful (↓)23.1%20.8%14.9%
Clean harmless (↑)97.4%97.4%97.2%

Under the same evaluation, BILD reaches 12.3% attacked harmful compliance. The roughly 47-point gap is therefore explained by the method — representation-level refusal-feature ablation versus behavior-level distillation — not by the training configuration.

W3. "The paper lacks sufficient implementation analysis. For example, why is a three-layer design necessary for TLCA?"

We added three ablation studies on Gemma-2-9B (attacked harmful compliance; higher = stronger attack):

AblationSettingResultReference
Layer count1 layer0.478Appendix K, Table 38
2 layers0.444
3 layers (full)0.856
Prefix–suffixPrefix only0.452Appendix J, Tbl 36–37
(avg. over 5 models)Suffix only0.350
Full (P+S)0.723
Length sensitivityPrefix 8–12 tokens0.824–0.887Appendix L, Tbl 39–40
Suffix 18–22 tokens0.832–0.869

The three-layer attack reaches the highest compliance in all fourteen languages individually, which motivates the design. Targeting multiple layers gives the attack objective coverage over refusal-related signals at different depths. Both prefix and suffix contribute, with architecture-dependent shares. Compliance varies only mildly across length settings, so TLCA does not rely on narrowly tuned length hyperparameters.

We hope these additions resolve the reviewer's concerns, and we are happy to run further analyses if useful.

Reviewer_rFjeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TLCA, a representation-guided attack that weakens refusal behavior in safety-aligned LLMs using ordinary textual prefixes/suffixes, and BILD, a LoRA-based defense that distills the model’s clean behavior to preserve robustness under attack.

Critical strengths
  1. The paper presents a coherent study of multilingual refusal robustness by proposing both a new attack (TLCA) and a corresponding defense (BILD).
  2. The paper evaluates five models across 14 languages, providing extensive multilingual results and showing consistent trends.
Critical weaknesses
  1. The baselines only include ReFAT, which is insufficient to fully demonstrate the effectiveness of the proposed method.
  2. BILD is trained and evaluated against a specific attack type (TLCA), so its robustness to other attack families remains unclear.
  3. The attack prefix and suffix are the same in both the training and test splits, raising the concern that the model may simply learn a specific pattern rather than genuinely improving refusal robustness on harmful requests.
Questions for authors

see waekness.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer

We thank the reviewer for recognizing the coherence of the attack–defense study and the extensive multilingual evaluation. We address each weakness below; where the concern points to an experiment we had not run, we have added it, and where it reflects a misunderstanding, we clarify the relevant design choice and point to the results already in the paper.

W1. "The baselines only include ReFAT, which is insufficient to fully demonstrate the effectiveness of the proposed method."

We added SafeDecoding (Xu et al., 2024) as a second defense baseline, implemented following the original paper and official code: α=3\alpha=3, m=2m=2 safety-aware decoding steps, c=5c=5 common candidate tokens, greedy decoding, and a LoRA refusal expert (rank 16, α=64\alpha=64) trained on the 36 English seed instructions with two accepted refusal completions each. All five models are evaluated across all fourteen languages (Table 2; implementation details in Appendix F).

Averaged over the five models (Table 2), BILD outperforms SafeDecoding on both axes simultaneously. Here "attacked harmless" means the TLCA prefix–suffix is wrapped around benign (harmless) test instructions, and we measure how often the model still answers them helpfully:

Metric (avg. over 5 models)SafeDecodingBILD
Harmful compliance, TLCA-attacked (↓)40.0%16.4%
Harmless compliance, TLCA prefix–suffix added (↑)68.4%96.0%

On InternVL3-8B, SafeDecoding retains only 28.8% of benign responses under attack. Under the same evaluation protocol, BILD achieves the best harmful-compliance and harmless-compliance results simultaneously.

W2. "BILD is trained and evaluated against a specific attack type (TLCA), so its robustness to other attack families remains unclear."

We added evaluations against two attack families that BILD never saw during training: a GCG-style universal optimized suffix (Appendix C) and an AutoDAN-style evolved readable template (Appendix D). Results are macro-averaged over fourteen languages (Table 1):

Attack (unseen in training)OriginalBILDΔ (pp)
GCG38.0%21.8%−16.2
AutoDAN44.9%18.8%−26.1

Δ\DeltaBILD is negative in every model–attack row of the table, and in all three attack conditions the defended harmful compliance is lower than the original model's clean harmful compliance (26.6%). We also note that TLCA remains the strongest of the three attacks on the original models (+45.7 pp on average, vs +11.4 pp for GCG and +18.3 pp for AutoDAN).

W3. "The attack prefix and suffix are the same in both the training and test splits, raising the concern that the model may simply learn a specific pattern."

Three observations address this concern directly.

  1. BILD cannot have learned a "wrapper →\rightarrow refuse" pattern, because its training mixture explicitly teaches it not to. Two of the four training components carry the identical TLCA prefix–suffix pair but require opposite behaviors (Section 3): adv-harmful examples (⟨\langleprefix$\rangle$ + harmful + ⟨\langlesuffix$\rangle$) target refusal, while adv-harmless examples (⟨\langleprefix$\rangle$ + harmless + ⟨\langlesuffix$\rangle$) target helpful answers. A pattern-matching adapter that fires on the attack wrapper would map both to refusal; BILD's objective instead rewards different outputs for the two given the same wrapper, forcing the decision onto the instruction content.

    The BILD rows of the comparison below are from the main evaluation; for completeness we have now additionally measured the original model under the same two conditions (Table 2). We take the benign (harmless) test instructions and measure how often the model still answers them helpfully, both with and without the TLCA prefix–suffix wrapped around them — four conditions in total:

    Input conditionModelHarmless compliance
    Clean (no attack)Original94.0%
    + TLCA prefix–suffixOriginal96.4%
    Clean (no attack)BILD96.0%
    + TLCA prefix–suffixBILD96.0%

    Macro-average over the five models and fourteen languages; higher is better.

    All four values lie within 2.4 points of each other, and BILD's attacked harmless rate (96.0%) is identical to its own clean rate and essentially unchanged from the original model's attacked rate (96.4%). If the adapter treated the attack wrapper as a refusal trigger, the last row would have collapsed toward 0% (as the Template LoRA-AT ablation demonstrates below). It does not: with the same prefix–suffix that decreases harmful compliance down to 16.4% on the BILD-trained model, BILD continues to answer benign prompts helpfully. The adapter is therefore conditioning on what the instruction asks, not on the presence of the attack string.

  2. The defense transfers to attack strings never seen during training (W2). GCG uses a gradient-optimized suffix only, and AutoDAN interleaves the instruction with an evolved readable template — both differ from the TLCA surface form. Yet BILD reduces harmful compliance under both (38.0% →\rightarrow 21.8% and 44.9% →\rightarrow 18.8%, Table 1) without having seen either attack. A rule keyed on the specific TLCA wrapper would not generalize to these structurally different attacks.

  3. The Template LoRA-AT ablation (Section 4.2, Appendix I) is precisely the "learn a specific pattern" condition, and it fails exactly as the reviewer predicts. When trained with a fixed refusal template, the adapter reaches 0% attacked harmful compliance — but harmless compliance also approaches 0% and generations collapse to a single repeated refusal string. BILD was designed to avoid exactly this failure mode via prompt-specific teacher targets, and the response-diversity statistics in Appendix I confirm that BILD generations remain prompt-specific rather than template-driven.

We also clarify the data split: the training and test instructions are disjoint (the official PolyRefuse train/test split); only the universal attack artifact (the prefix–suffix pair) is shared between train and test.

Ethics flag. We have added an Ethics Statement (page 51). The work is defensive research: the paper reports only aggregate results and contains no optimized attack strings or harmful model generations; any release of attack artifacts will follow controlled-access and responsible-disclosure practices; all experiments were conducted in controlled academic environments on publicly available, license-compliant datasets. We are happy to expand this statement if the chairs consider it necessary.

AuthorsOpenReview ↗

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer rFje,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.